Re: [Qemu-devel] [QEMU-devel][PATCH v3] aio-posix: Fix concurrent aio_poll/set_fd_handler.

[Remy NOEL]

On 12/18/18 6:39 PM, Paolo Bonzini wrote:

> On 17/12/18 17:48, address@hidden wrote:
> > Also, we do not call aio_epoll_update for deleted handlers as this has
> > no impact whatsoever.
> Why?  epoll is used in level-triggered mode, so you do have to remove
> the file descriptor...
>
> >                  atomic_read(&ctx->poll_disable_cnt) + poll_disable_change);
> >   
> > -    aio_epoll_update(ctx, node, is_new);
> > +    if (new_node) {
> > +        aio_epoll_update(ctx, new_node, is_new);
> > +    }
> >       qemu_lockcnt_unlock(&ctx->list_lock);
> >       aio_notify(ctx);
> ... so I think this should be "if (node || new_node)"?

Well, currently, when an AioHandler is removed, we do not change 
node->pdf.events (only revents).

Therefore a call to aio_epoll_update on node will only result in a call 
to epoll_ctl with EPOLL_CTL_MOD and the same event, which seems kinda 
pointless.

we may set node->pfd.events to 0 to unregister the file descriptor, but 
this would change the behavior compared to current handling of node 
deletion if i'm not mistaken.

Remy